In short. Multi-hop question answering combines evidence across documents to answer a question, with supporting facts that can be checked at each step.

Definition: Multi-hop question answering is answering a question by combining evidence from multiple documents. A reasoning step connects information across those documents, for example by identifying a person in one passage and finding a fact about that person in another.

Key facts

The HotpotQA paper, published in 2018, introduces a Wikipedia-based question-answering dataset with sentence-level annotations identifying the evidence needed for answers.
The 2023 IRCoT paper describes a retrieval method in which intermediate reasoning guides further searches, and the retrieved text informs subsequent reasoning.
The 2022 MuSiQue paper constructs questions with connected reasoning steps and adds unanswerable variants to test whether the necessary evidence is available.

What does a hop look like?

A bridge entity is the person, place or other entity that connects passages. The HotpotQA paper illustrates this with a question about the birthday of Radiohead's singer and songwriter. The reader first identifies Thom Yorke, then looks for his birthday. Thom Yorke is the bridge entity. The following breakdown follows that example:

Question: when was the singer and songwriter of Radiohead born?
Intermediate entity: Thom Yorke
Next fact to find: Thom Yorke's birthday

The paper also includes comparison questions. These require extracting properties of different entities and comparing them, such as determining whether bands come from the same country. A multi-hop question can therefore involve either following a connection or bringing separate facts together for a comparison.

Simply collecting several documents creates a different requirement from combining their evidence. HotpotQA's introduction describes earlier datasets that supply multiple documents even though the answer often appears in nearby sentences within a single paragraph.

Why can retrieval require several steps?

Retrieval is the selection of potentially relevant text from a collection. The IRCoT paper explains a dependency in multi-step questions: the information to retrieve can depend on an intermediate result, which itself depends on previously retrieved material. Its proposed method alternates retrieval with generated reasoning sentences. That makes an intermediate result useful for deciding what evidence to seek next.

The HotpotQA paper separates reading from the additional difficulty of finding evidence through its benchmark settings. In the distractor setting, the system receives the two supporting paragraphs mixed with eight other paragraphs. In the full wiki setting, it must locate evidence among the opening paragraphs of Wikipedia articles, without the supporting paragraphs being identified in advance.

These settings distinguish a failure to find the relevant material from a failure to combine material already supplied. For website reviews, an application of this distinction is to check separately whether the relevant page is retrieved and whether its passage supports the required relationship.

How can the evidence chain be checked?

Supporting facts are the sentences needed to justify an answer. The HotpotQA paper collects these sentences from annotators and evaluates predicted supporting sentences against the annotated set. It also defines joint evaluation measures combining answer correctness with supporting-fact performance.

The MuSiQue paper addresses shortcuts in benchmark questions. Its construction method connects single-hop questions so that a reasoning step depends on information from another step. MuSiQue-Full adds unanswerable contrast questions. This gives an evidence review a concrete focus: identify the fact each step needs, then check whether the supplied passages contain it and connect to the next step.

FAQ

Is every question with several retrieved pages multi-hop?

A multi-hop question requires combining evidence across documents. The HotpotQA paper distinguishes this from datasets where several documents are supplied but a single paragraph can answer the question.

How can I review an answer that combines pages?

You can identify the supporting sentence for each required fact and check the connection between them. This applies the sentence-level evidence approach in the HotpotQA paper, which evaluates supporting facts alongside answers.

Sources

Check how a page reads to answer engines. Open the GEO score analyzer